Clinical Prediction Models Often Lack Shared Code
A review of nearly 4,000 studies found only about 1 in 8 reported sharing the code behind their prediction models.
Updated on Nov. 19, 2026 in Cancer

An international team led by MIT and Dana-Farber Cancer Institute researchers used large language models to examine nearly 4,000 clinical prediction studies and hundreds of code repositories. The findings were published in Nature Medicine and are informing TRIPOD-Code, a reporting guideline under development.
Why it matters
Clinical prediction models help guide patient care, so scientists need to be able to scrutinize how they were built and tested. Making code public does not always mean it includes the documentation needed for others to reuse it.
The review covered nearly 4,000 studies with full texts in PubMed Central's open-access collection, and about 1 in 8 reported sharing code, with rates across journals ranging from zero to more than 70 percent. Which elements of shared code should be considered essential for reuse has not yet been settled.
The players
Massachusetts Institute of Technology (MIT)
MIT is a private research university in Cambridge, Massachusetts, known for its leading work in engineering, computer science, and artificial intelligence.
Dana-Farber Cancer Institute
Dana-Farber Cancer Institute is a Boston-based cancer research and treatment center affiliated with Harvard Medical School.
The details
Researchers used an LLM to screen papers and identify statements about code availability, then software retrieved linked repositories and assessed documentation completeness, software requirements, and terms of reuse. Some studies shared synthetic datasets where original patient records could not be released publicly.
Timeline
Findings were published on MIT News on November 19, 2026, alongside the Nature Medicine paper.
Big Picture
The work builds on the TRIPOD reporting guideline for prediction model studies by proposing TRIPOD-Code to define the analytical code that should accompany published models. It extends existing reporting standards from describing methods in text to making the underlying software reusable and scrutinizable.
Clinical prediction tools can influence decisions about treatment and monitoring, so greater transparency about how they were built can help clinicians and patients better understand their limits. Shared, reusable code makes it easier for other teams to test whether a model performs as claimed in different patient populations.
The takeaway
Readers evaluating health tools should look for whether a prediction model's code and documentation are publicly available, since sharing code alone may not allow others to reuse it. Greater transparency helps determine whether a tool's reported accuracy holds up beyond the original study.
Further reading
Explore more on Cancer research and the science behind how new diagnostic and prognostic tools are evaluated.
Source note: This article includes information reported by MIT News | Massachusetts Institute of Technology.



